Compute Subcellular Spatial Features

Author: Clarence Mah | Last Updated: 6/15/2022

Here we will demonstrate how to compute various subcellular spatial features with bento. We will use the included seqFISH+ dataset.

Load Libraries

import bento
import seaborn as sns

Load Data

Let’s start with the preprocessed seqFISH+ data.

adata = bento.datasets.load_dataset('seqfish')

Cell Features

There are a number of tools to compute cell-level features. List available features like so.

bento.tl.cell_features.keys()

dict_keys(['cell_span', 'cell_bounds', 'cell_moments', 'raster_cell', 'cell_aspect_ratio', 'cell_density', 'cell_area', 'cell_perimeter', 'cell_radius'])

We can compute a single feature…

bento.tl.cell_area(adata)

AnnData object modified:
    obs:
        + cell_area

Or conveniently, a whole set of features simultaneously.

some_features = ["cell_area", "cell_density", "cell_aspect_ratio"]
bento.tl.analyze_cells(adata, some_features)

AnnData object modified:
    obs:
        + cell_density, cell_aspect_ratio

For starters, we can study pairwise relationships of these cell features.

sns.pairplot(
    data=adata.obs[["cell_area", "cell_density", "cell_aspect_ratio"]],
    kind="reg",
)

<seaborn.axisgrid.PairGrid at 0x1554e0dc0fd0>

sns.clustermap(
    adata.obs[["cell_area", "cell_density", "cell_aspect_ratio"]].T,
    standard_scale=0,
    figsize=(8, 3),
    cmap="Reds",
)

<seaborn.matrix.ClusterGrid at 0x1554c7f9a2e0>

dense_cells = adata.obs["cell_density"].sort_values(ascending=False).index[:10]
bento.pl.cellplot(adata[dense_cells], col="cell", col_wrap=5, height=3)

Sample Features

Similarly, we can compute subcellular spatial features for samples (sample corresponds to molecules for a given cell-gene pair), to produce high-dimensional feature spaces.

bento.tl.sample_features.keys()

dict_keys(['cell_proximity', 'nucleus_proximity', 'cell_asymmetry', 'nucleus_asymmetry', 'point_dispersion', 'nucleus_dispersion', 'ripley_stats', 'nucleus_enrichment'])

bento.tl.analyze_samples(
    adata, ['cell_proximity', 'nucleus_proximity', 'point_dispersion'], chunksize=100
)

Note

Bento is able to utilize these feature spaces for predicting RNA localization patterns. For unsupervised analysis, Bento implements tensor decomposition to generate interpretable low dimensional signatures.